JPS5872995A - Word voice recognition - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a word speech recognition method that first performs phoneme recognition on input speech and then recognizes words by comparing this recognized phoneme sequence with a word dictionary in which phonemes are expressed. The purpose of this method is to improve the word recognition rate by distinguishing between voiced vowels and voiced vowels.

First, a conventional word speech recognition method of this type will be explained with reference to FIG.

As shown in Figure 1, the input word speech is analyzed, the features of this input word speech are extracted, the phonemes that make up the input word speech are recognized, and the recognized phoneme sequence is stored in a word dictionary with phoneme notation. Each word and the conflation matrix (Co
(hereinafter abbreviated as C, M)
, the likelihood is calculated, and the word with the highest likelihood is selected as the recognized word.

Table 1 shows an example of the word dictionary (city names) expressed in phonemes.

Table 1 Table 2 also shows an example of phoneme notation in the word dictionary.

As shown in Table 1, the phoneme notation in the word dictionary is as if a
-~ It was done mechanically, as if writing letters, and no consideration was given to devoicing °. Furthermore, in the conventional example, phoneme recognition is performed at the phoneme recognition stage (Figure 1) based on physical parameters such as the spectrum of speech, and as a result, when a voiceless consonant with an extremely long duration is recognized, teeth,
The consonant was regarded as a series of three phonemes: an unvoiced consonant phoneme, a devoiced vowel phoneme, and an unvoiced consonant phoneme, and the phonetic sequence was modified. For example, if C of the long-duration voiceless consonant CIBU is recognized, three phonemes CUI
C$ CUICIBU considers the audio sequence to be continuous.
It was corrected. Which vowels are usually devoiced? or U, so the vowel added in the modification is the intermediate vowel UI between 葡 and U.

FIG. 2 shows part of C and M in the conventional example.

The numbers in C and M are the probability P (W/D
) is expressed in chi. Recognized phoneme U in C, M.
I has a voiced ubiquitous tone whose physical properties are intermediate between U and I, and
It has 1 duration, 1 duration, and 9 voiceless consonants and a modified devoiced vowel as a result of recognition.

In Japanese, as a general rule, consonants other than ``n'' are always followed by a voiced vowel or semi-vowel. For example, NARA (
Oak) K J OOT O (Kyoto) etc. As long as it is based on this principle, it will not be recognized by the conventional method. Furthermore, even in the conventional example, countermeasures for devoicing have been taken through modification.

However, in reality, ■ and U in words are devoiced by about 20 characters. For this reason, in the conventional example, countermeasures against devoicing were insufficient, resulting in the drawback of erroneous recognition of words.

The present invention eliminates the drawbacks of the above conventional example,
An embodiment of the present invention will be described below.

In the word dictionary of this embodiment, vowels that are likely to be devoiced are given a devoicing symbol in advance, and are given a different phoneme notation from the original voiced vowel. For example, the city name Fuchu ``Fuchiku'' is written in conventional word dictionaries as shown in Table 1 (HUCJUU
), but since "fu" is easily devoiced, in this example, it is expressed as ()IU-CJIJU ), and U
- and U are different phonemes.

Furthermore, in this embodiment, the vowel added to correct the phoneme sequence as a measure against devoicing in the phoneme recognition stage is expressed as UI-, to distinguish it from voiced UI.

FIG. 3 shows part of C and M in this embodiment. In Figure 3, the phonemes I, I-tU and U in the word dictionary
Comparing how each - was recognized,
It can be seen that voiced sounds and unvoiced sounds are clearly separated. Also, from Figure 3, it can be seen that vowels that are easily devoiced are easily dropped or dropped during phoneme recognition. 1. Shibashiba (SUISEU) Totosu, in this case the word recognition result is "Shinshiku" instead of "Fuchikou"
(In the word dictionary, it was (SIN=ZJUKU)). On the other hand, according to this embodiment, the phoneme recognition result of the same data as the conventional example for "Fuchu" (tHu-c Juu in the word dictionary) is (SUI-8EU).
The word recognition result was correctly ``fuchu''. This is because while the conventional recognized phoneme (Ul) had the same likelihood for both U and ■ in the word dictionary, the recognized phoneme (UI-) in this embodiment This is because (U-) in ``Shinnok'' has a high likelihood, and +I) in ``Shinnok'' has only a low likelihood.

In the above embodiment, I- is used as a phoneme in the word dictionary.

Although only U- is added, it may be further divided depending on the frequency of vowel devoicing. Although it is rare, E
, A, and O may also be devoiced, and these devoiced vowels may be added. In addition, in the above embodiment, UI is used as a recognized phoneme.
Although only - is added, if it becomes possible to separate U-2- and then add U- and I- as recognized phonemes, the word recognition rate can be further improved.

As described above, the present invention has the advantage that the word recognition rate can be improved because the devoiced vowels are distinguished from the original voiced vowels.

[Brief explanation of the drawing]

Fig. 1 shows an outline of the word speech recognition method. Fig. 2 shows a part of C and M in the conventional example. Fig. 3 shows the C used in the word speech recognition method in an embodiment of the present invention. ,M,
FIG. Figure 1

Claims (2)

[Claims] (1) Perform phoneme recognition on the input speech to obtain a recognized phoneme sequence, and use a confusion matrix to calculate the likelihood between this recognized phoneme and a phoneme in a word dictionary with phoneme notation to recognize words. In the word speech recognition method, the phoneme symbols representing vowels in the word dictionary are written separately into normally generated voiced vowels and vowels that are easily devoiced, and the above-mentioned phoneme symbols are represented as dictionary phoneme items in the confusion matrix. A word speech recognition method 1 characterized in that voiced vowels and vowels that are easily devoiced are provided separately. (2) In the word speech recognition method, voiced vowels and unvoiced vowels are distinguished as the types of recognized phonemes that are the result of phoneme recognition, and the recognized phoneme items of the conflation matrix are adjusted to the types of recognized phonemes. The word speech recognition method according to claim 1, wherein voiced vowels and voiceless vowels are provided separately.